\r \r

MCP Tool Integrator\r

\r

Overview\r

\r The Model Context Protocol (MCP) is the emerging standard for connecting AI agents to external tools — and 2026 is the year it goes mainstream. Microsoft Agent 365, Claude Desktop, Cursor, and dozens of frameworks now support MCP natively. This skill helps developers scaffold, configure, and debug MCP server integrations at speed — turning scattered tool APIs into a unified agent capability layer.\r \r

Title\r

\r MCP Tool Integrator — Connect AI Agents to Any Tool in Minutes\r \r

Triggers\r

\r

• "MCP server setup" / "MCP服务器配置"\r
• "MCP integration" / "MCP集成" / "MCP接入"\r
• "Model Context Protocol" / "MCP协议"\r
• "Claude MCP tools" / "Claude MCP工具"\r
• "AI agent tool integration" / "AI代理工具集成"\r
• "MCP GitHub" / "MCP Slack" / "MCP Notion"\r
• "MCP debug" / "MCP调试"\r
• "MCP LangChain" / "MCP n8n"\r
• "build MCP server" / "构建MCP服务器"\r
• "MCP custom tool" / "MCP自定义工具"\r \r ---\r \r

0. 2025-2026 MCP 最新动态\r

\r | 时间 | 动态 | 意义 |\r |------|------|------|\r | 2024年11月 | Anthropic发布MCP协议 | AI Agent与外部工具交互的开放标准诞生 |\r | 2025年11月 | MCP移交至Linux Foundation旗下Agentic AI Foundation治理 | OpenAI、Google、Microsoft等主要厂商共同参与标准制定 |\r | 2026年 | MCP成为AI Agent开发事实标准协议 | Microsoft Agent 365、Claude Desktop、Cursor等主流平台原生支持 |\r | 2026年 | FastMCP简化MCP Server开发 | 开发者可快速搭建自定义MCP服务器 |\r | 2026年5月 | MCP官方Server Registry扩展至50+工具 | 涵盖GitHub、Slack、Notion、Postgres、腾讯云等 |\r \r

关键提示： 2026年MCP生态已从单一AI厂商协议演变为跨平台开放标准。金融行业部署MCP时，优先使用官方认证的Server；国内企业可选用国产MCP Server（如腾讯云、钉钉、飞书定制实现）。\r \r ---\r \r

Workflow\r

\r

Phase 1 — MCP Fundamentals & Environment Setup\r

\r Step 1.1: Detect Current MCP Environment\r \r Determine what MCP runtime is available and what tools are already connected.\r \r Output: MCP Environment Audit\r \r | Runtime | Version | Connected Tools | Status |\r |---------|---------|----------------|--------|\r | Claude Desktop MCP | 1.0.3 | filesystem, github | ✅ Active |\r | Cursor MCP Bridge | 0.9.2 | postgres, slack | ⚠️ Partial |\r | Custom LangChain MCP | N/A | None | ❌ Not configured |\r \r Step 1.2: Recommend MCP Architecture\r \r Based on use case, recommend the optimal MCP topology.\r \r Architecture Patterns:\r \r

Pattern A — Desktop-First (Individual Developer)\r
Claude Desktop ↔ Local MCP Servers ↔ filesystem, git, terminal\r
\r
Pattern B — Enterprise Multi-Agent (Team)\r
LangChain Agent ↔ MCP Gateway ↔ GitHub, Jira, Slack, Notion, Postgres\r
\r
Pattern C — API-First (Production)\r
FastAPI MCP Server ↔ Authenticated Tools ↔ CRM, ERP, Database\r
\r
Pattern D — China-Optimized (Regulated Industry)\r
Local MCP Server ↔ Domestic tools (钉钉, 飞书, 腾讯云) ↔ Firewall-compliant\r
```\r
\r
---\r
\r
### Phase 2 — Scaffolding MCP Servers\r
\r
**Step 2.1: Generate MCP Server Code**\r
\r
For any external tool, generate a complete MCP server implementation.\r
\r
**Input:** Tool name + authentication method + required operations\r
**Output:** Complete Python/TypeScript MCP server scaffold\r
\r
**Example MCP Server — Notion Integration:**\r
\r
```python\r
# notion_mcp_server.py\r
from mcp.server.fastapi import McpServer\r
from mcp.types import Tool, CallToolRequest\r
import httpx\r
\r
SERVER = McpServer(name="notion-mcp", version="1.0.0")\r
\r
@SERVER.list_tools()\r
async def list_notion_tools():\r
    return [\r
        Tool(\r
            name="notion_search_pages",\r
            description="Search Notion pages by keyword",\r
            input_schema={\r
                "type": "object",\r
                "properties": {\r
                    "query": {"type": "string"},\r
                    "filter_database_id": {"type": "string", "optional": True}\r
                }\r
            }\r
        ),\r
        Tool(\r
            name="notion_create_page",\r
            description="Create a new Notion page in a database",\r
            input_schema={\r
                "type": "object",\r
                "properties": {\r
                    "database_id": {"type": "string"},\r
                    "title": {"type": "string"},\r
                    "properties": {"type": "object"}\r
                }\r
            }\r
        ),\r
        Tool(\r
            name="notion_update_block",\r
            description="Update a block in a Notion page",\r
            input_schema={\r
                "type": "object",\r
                "properties": {\r
                    "block_id": {"type": "string"},\r
                    "content": {"type": "string"}\r
                }\r
            }\r
        )\r
    ]\r
\r
@SERVER.call_tool()\r
async def call_notion_tool(request: CallToolRequest):\r
    if request.name == "notion_search_pages":\r
        return await search_pages(request.arguments["query"], request.arguments.get("filter_database_id"))\r
    elif request.name == "notion_create_page":\r
        return await create_page(request.arguments["database_id"], request.arguments["title"], request.arguments.get("properties", {}))\r
    elif request.name == "notion_update_block":\r
        return await update_block(request.arguments["block_id"], request.arguments["content"])\r
```\r
\r
**Step 2.2: MCP Server Configuration File**\r
\r
Generate the `mcp.json` or `mcp_servers.json` config for the runtime.\r
\r
```json\r
// .mcp.json (Claude Desktop)\r
{\r
  "mcpServers": {\r
    "notion": {\r
      "command": "python",\r
      "args": ["notion_mcp_server.py"],\r
      "env": {\r
        "NOTION_API_KEY": "${NOTION_API_KEY}"\r
      }\r
    },\r
    "github": {\r
      "command": "npx",\r
      "args": ["-y", "@modelcontextprotocol/server-github"],\r
      "env": {\r
        "GITHUB_TOKEN": "${GITHUB_TOKEN}"\r
      }\r
    },\r
    "postgres": {\r
      "command": "npx",\r
      "args": ["-y", "@modelcontextprotocol/server-postgres"],\r
      "env": {\r
        "DATABASE_URL": "${DATABASE_URL}"\r
      }\r
    }\r
  }\r
}\r
```\r
\r
---\r
\r
### Phase 3 — Connecting Popular Tool Chains\r
\r
**Step 3.1: GitHub MCP Integration**\r
\r
Enable AI agents to interact with GitHub repositories, issues, PRs, and code.\r
\r
**Capabilities enabled:**\r
- `github_list_repos` — List repositories with filters\r
- `github_create_issue` — Create issue with labels\r
- `github_review_pr` — Analyze PR changes and provide review comments\r
- `github_search_code` — Semantic code search across repos\r
- `github_get_workflow_runs` — Monitor CI/CD pipeline status\r
\r
**Use case example:**\r
> "Summarize all open PRs in our main repo, highlight security concerns, and post a daily digest to Slack."\r
\r
**Step 3.2: Slack MCP Integration**\r
\r
Enable AI agents to send messages, search history, manage channels.\r
\r
**Capabilities enabled:**\r
- `slack_post_message` — Send to channel or DM\r
- `slack_search_messages` — Full-text search in Slack history\r
- `slack_list_channels` — Get channel list with membership\r
- `slack_create_channel` — Provision new channels\r
- `slack_schedule_message` — Schedule future messages\r
\r
**Step 3.3: Database MCP (PostgreSQL / MySQL / MongoDB)**\r
\r
Enable AI agents to query databases, generate reports, and validate data.\r
\r
**Capabilities enabled:**\r
- `db_query` — Execute read-only SQL with row limits\r
- `db_describe_table` — Get schema documentation\r
- `db_generate_report` — Natural language → formatted report\r
- `db_validate` — Check data quality rules\r
\r
**⚠️ Security note:** Always use read-only connections. Never expose write permissions without approval workflows.\r
\r
---\r
\r
### Phase 4 — Debugging & Optimization\r
\r
**Step 4.1: MCP Connection Diagnostic**\r
\r
When an MCP tool fails, systematically diagnose the root cause.\r
\r
**Diagnostic checklist:**\r
1. Authentication — Is the API key valid and not expired?\r
2. Network — Can the server reach the tool's API endpoint?\r
3. Permission — Does the token have the required scopes?\r
4. Rate limit — Is the tool's API rate limit exceeded?\r
5. Schema mismatch — Does the tool's input_schema match the server's definition?\r
6. Runtime compatibility — Is the MCP server version compatible with the runtime?\r
\r
**Output: MCP Debug Report**\r
\r
| Tool | Connection Status | Latency | Last Error | Root Cause |\r
|------|-----------------|---------|-----------|------------|\r
| Notion | ❌ Failed | — | 401 Unauthorized | Token expired |\r
| GitHub | ✅ Working | 230ms | — | — |\r
| Postgres | ⚠️ Slow | 4500ms | — | Missing index on query |\r
\r
**Step 4.2: Performance Optimization**\r
\r
For slow MCP tools, recommend optimizations:\r
- Add database indexes for common query patterns\r
- Implement response caching (TTL: 5 minutes for read-only tools)\r
- Batch API calls where the tool supports it\r
- Use streaming responses for large data\r
\r
---\r
\r
## Input / Output Examples\r
\r
### Example 1: Building a Multi-Tool Agent Chain\r
\r
**Input:**\r
```\r
I want to build an agent that:\r
1. Reads feature requests from Notion\r
2. Creates GitHub issues for each one\r
3. Posts a summary to Slack\r
4. Updates Notion with the GitHub issue links\r
\r
I use Claude Desktop. Help me set up the MCP servers.\r
```\r
\r
**Output:**\r
```\r
## MCP Architecture for Feature Request Agent\r
\r
### Required MCP Servers (3 total)\r
1. notion-mcp — Read/write Notion databases\r
2. github-mcp — Create GitHub issues\r
3. slack-mcp — Post Slack notifications\r
\r
### Setup Instructions\r
\r
Step 1: Install MCP servers\r
```bash\r
npm install -g @modelcontextprotocol/server-notion\r
npm install -g @modelcontextprotocol/server-github  \r
npm install -g @modelcontextprotocol/server-slack\r
```\r
\r
Step 2: Configure .mcp.json in Claude Desktop config directory\r
\r
Step 3: Verify connections with diagnostic check\r
\r
### Agent Prompt Template\r
"You are a feature request manager. Every morning at 9am:\r
1. Search Notion for items in the 'Feature Requests' database with status='pending'\r
2. For each item, create a GitHub issue in repo 'acme/product' with label 'feature-request'\r
3. Reply to each Notion item with the GitHub issue URL\r
4. Post a Slack message to #product-updates with count + top 3 priority items"\r
\r
### Error Handling\r
- If GitHub API rate limited: retry after 60s, queue remaining items\r
- If Notion update fails: log to error Notion page, alert via Slack DM\r
```\r
\r
### Example 2: Debugging a Failing MCP Connection\r
\r
**Input:**\r
```\r
My Slack MCP integration stopped working. The error says \r
"channel_not_found" when I try to post messages.\r
```\r
\r
**Output:**\r
```\r
## MCP Slack Debug Report\r
\r
### Diagnostic Results\r
1. Authentication: ✅ Token valid (expires in 45 days)\r
2. Network: ✅ API endpoint reachable (45ms)\r
3. Permission: ⚠️ Token missing scopes: channels:write\r
4. Rate Limit: ✅ Well within limits (12% used)\r
5. Schema: ✅ Input schema matches server definition\r
6. Runtime: ✅ Version 1.0.3 compatible\r
\r
### Root Cause Identified\r
MISSING SCOPE: The Slack app's OAuth token does not include\r
the "channels:write" scope required to post to public channels.\r
\r
### Fix Instructions\r
1. Go to api.slack.com/apps → your app → OAuth & Permissions\r
2. Add scope: "channels:write" (or "chat:write" for broader access)\r
3. Reinstall the app to workspace (OAuth reinstall required for scope changes)\r
4. Update your MCP config with the new token\r
5. Re-run the diagnostic to confirm ✅\r
\r
### Alternative Workaround\r
Use the conversations.open API instead — it works with existing scopes\r
and can post to any channel the bot has been invited to.\r
```\r
\r
---\r
\r
## MCP Server Registry\r
\r
Pre-built MCP server templates available:\r
\r
| Tool | Package | Auth | Operations | China Status |\r
|------|---------|------|-----------|-------------|\r
| GitHub | @modelcontextprotocol/server-github | OAuth | 20+ | ✅ Works globally |\r
| Slack | @modelcontextprotocol/server-slack | OAuth | 15+ | ⚠️ Slack blocked in China |\r
| Notion | @modelcontextprotocol/server-notion | API Key | 12+ | ✅ Works globally |\r
| PostgreSQL | @modelcontextprotocol/server-postgres | Connection string | 5+ | ✅ Works globally |\r
| Filesystem | Built-in | Local | 8+ | ✅ Works globally |\r
| Brave Search | @modelcontextprotocol/server-brave-search | API Key | 3+ | ⚠️ Limited |\r
| AWS | aws-mcp | AWS credentials | 30+ | ✅ S3/lambda work |\r
| 腾讯云 | Custom (not official) | SecretKey | Varies | ✅ China-optimized |\r
\r
---\r
\r
## Notes & Best Practices\r
\r
1. **MCP vs. direct API calls:** MCP adds a layer of standardization. Use it when you need to swap AI runtimes (Claude ↔ GPT ↔ Gemini) without rewriting tool integrations.\r
2. **China-specific:** Official MCP servers for 钉钉 (DingTalk) and 飞书 (Lark) are not in the official registry — build custom ones using the MCP Python SDK. 腾讯云 SDK has partial MCP compatibility.\r
3. **Security:** MCP tools inherit the AI agent's access level. Always use least-privilege tokens and enable audit logging.\r
4. **Versioning:** MCP moved to Linux Foundation Agentic AI Foundation (2025-11). Pin server versions in production. For China deployments, track domestic MCP ecosystem evolution.\r
5. **Testing:** Use `mcp dev` CLI or the Claude Desktop MCP inspector to test tools before deploying to agents.\r
6. **Cost control:** Many MCP tool calls count as API calls. Set rate limits and budgets per agent.\r
7. **2026标准之战：** MCP vs. OpenAI Tool Use vs. Google A2A — MCP已获得最多生态支持，但跨协议互操作性是2026年新挑战。使用标准转换层（如MCP Gateway）可桥接不同协议。\r
8. **企业AI Agent首选：** 金融行业部署AI Agent时，优先通过MCP接入内部系统（CRM/ERP），而非直接API集成——MCP的审计日志和访问控制更规范。\r
\r
---\r
\r
*Author: @gechengling | Skill: mcp-tool-integrator | clawhub.ai/gechengling/mcp-tool-integrator*\r